基於深度學習捲積神經網路之電腦斷層掃描肺腺癌肋膜侵犯預測模型

Prediction Model of Visceral Pleural Invasion in Lung Adenocarcinoma on Computed Tomography Based on Deep Learning Convolutional Neural Network

汪傳崴(Chuan-Wei Wang)

指導教授 ： 陳中明

國立臺灣大學/工學院/醫學工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202200332

摘要

根據民國108年衛福部統計，癌症已連續38年位居國人十大死因之首，其中又以肺癌為兩性癌症死亡率之首，透過電腦斷層掃描(CT)之肺癌篩檢，能夠及早發現早期肺癌，進而及早進行治療，提高肺癌存活率。在早期肺癌之手術治療策略中，除了腫瘤大小外，腫瘤有無肋膜侵犯亦會影響肺癌手術術式與切除範圍。肋膜侵犯被認為是不良的預後因子，有較低的五年存活率與較高的局部復發機率，因此有肋膜侵犯的患者較不建議採取手術範圍較小之次肺葉切除術，肺葉切除術仍為此類患者的優先選擇。
臨床上判別腫瘤是否有肋膜侵犯，主要依靠CT影像上之影像特徵與臨床醫師之先驗知識，根據臨床醫師經驗不同，於判斷肋膜侵犯時會存在觀察者間之差異，且可能因疲勞增加誤判的可能性，因此，開發電腦輔助診斷系統將有助於幫助醫師判別有無肋膜侵犯，可以提供臨床醫師進行影像判斷時的第二種意見，降低觀察者間的差異，提高臨床診斷的性能與增加分類的準確性。
本研究為輔助臨床醫師進行腫瘤肋膜侵犯之判別，提出(1)基於捲積神經網路之深度學習演算法；(2)加入注意力機制於深度學習網路中，以期達到肋膜侵犯判別之理想結果，盡可能地區分出肋膜侵犯患者，擁有高靈敏度的同時，擁有良好的特異度，最大化避免非肋膜侵犯患者接受不必要之外科手術，並使深度學習網路更加專注於欲觀察之目標上。
本研究於基於捲積神經網路之深度學習演算法中，提出4 Layers Convolutional Neural Network(4L CNN)之網路架構，進行10-fold交叉驗證，可得此模型之Accuracy為0.764±0.046，Sensitivity為0.691±0.173，Specificity為0.776±0.057，於Sensitivity之表現較差。以積分梯度將深度學習模型可視化，模型之訓練結果較多關注於腫瘤邊緣、腫瘤與肺壁接觸部分等，但容易有模型關注部分於整個肺區以及肺壁部分之情況。
本研究於加入注意力機制於深度學習網路中，提出三種注意力機制: (1)Squeeze and Excitation block，(2)Dilate convolution block，(3)Lung map segmentation block，此三種注意力機制相較於單純使用4L CNN，Sensitivity皆有所進步，其中以加入Lung map segmentation block之10-fold結果及AUC表現最佳，可得模型Accuracy為0.778±0.028，Sensitivity為0.779±0.092，Specificity為0.778±0.030，AUC為0.8284。以積分梯度進行深度學習模型可視化，加入Squeeze and Excitation block之模型較容易有模型專注於整個VOI之情況，加入Dilate convolution block之模型能夠主要關注於腫瘤邊緣、肋膜標籤、肺區與肺壁之交界等，加入Lung map segmentation block之模型能夠主要關注於整體腫瘤、腫瘤邊緣、肋膜標籤、腫瘤或肺區與肺壁之交界等。
本研究使用添加Lung map segmentation block之深度學習網路架構測試外部資料，並與臨床醫師判別結果進行比較，臨床醫師之判別結果傾向Sensitivity較高，但Accuracy與Specificity較低，深度學習模型30次測試結果可得到Accuracy為0.822±0.030，Sensitivity為0.784±0.052，Specificity為0.828±0.037，雖於Sensitivity略低於臨床醫師，但Accuracy與Specificity皆高出許多。
本研究進一步將外部資料分為腫瘤有無接觸肋膜之情況，在腫瘤有接觸肋膜之情況下，深度學習模型雖於有肋膜侵犯判別之準確率略低於醫師判讀，但在無肋膜侵犯之判別高於醫師判讀，此外，醫師判讀結果落於深度學習模型ROC curve之95%信賴區間中，由實驗結果顯示，本研究提出之深度學習肋膜侵犯判別模型能與臨床醫師有相近之判別能力。

關鍵字

肺腫瘤肋膜侵犯分類 ； 深度學習 ； 捲積神經網路 ； 注意力機制 ； 深度學習模型可視化

並列摘要

According to the 2019 report of Ministry of Health and Welfare, cancer has been ranked as the first among the top ten causes of death in Taiwan for 38 consecutive years, and lung cancer is the leading cause of cancer mortality in both sexes. Having lung cancer screening through computed tomography(CT) can detect early stage lung cancer early, and have treatment early to improve the survival rate of lung cancer. In the surgical treatment strategy for early stage lung cancer, in addition to tumor size, the presence of visceral pleural invasion also affects surgical decision and extent of resection. Visceral pleural invasion is considered to be a poor prognostic factor, with lower five-year survival rate and higher chance of local recurrence. Therefore, patients with visceral pleural invasion are less recommended to undergo sublobar resection with a smaller surgical region. Lobectomy is still the preferred option for such patients with visceral pleural invasion.
To determine whether a tumor has visceral pleural invasion clinically, it mainly depends on the features on CT images and prior knowledge of clinicians. According to the experience of clinicians, there will be inter-observer differences in determining visceral pleural invasion, and the probability of misjudgment may be increased due to fatigue. Therefore, the development of computer-aided diagnosis system will help clinicians to determine visceral pleural invasion, which can provide a second opinion when making image judgements, reduce inter-observer differences, improve the performance of clinical diagnosis and increase classification accuracy.
In order to assist clinicians in diagnosing visceral pleural invasion, this study proposes:(1) a deep learning algorithm based on convolutional neural network, (2) adding attention mechanism to the deep learning network, to achieve the ideal visceral pleural invasion judgement. Patients with visceral pleural invasion can be distinguished as many as possible by the model, with high sensitivity and good specificity, maximizing the avoidance of unnecessary surgery for those without visceral pleural invasion, and making the deep learning network mainly focus on the target to be observed.
In this study, the network architecture of 4 Layers Convolutional Neural Network is proposed in the deep learning algorithm based on convolutional neural network. The model can achieve accuracy 0.764±0.046, sensitivity 0.691±0.173, specificity 0.776±0.057 through 10-fold cross validation, which shows poor performance in sensitivity. Using Integrated gradients to visualize the deep learning model, the results of the model are mainly focused on tumor margins, contact parts between tumor and lung wall, etc., but it is likely to focus on the entire lung area and the lung wall.
In this study, three attention mechanisms are proposed in adding attention mechanism to the deep learning network: (1)Squeeze and Excitation block, (2)Dilate convolution block, (3)Lung map segmentation block. Sensitivity of adding these three attention mechanisms are higher than the network using 4L CNN alone. Among these three attention mechanisms, the 10-fold performance and AUC of Lung map segmentation block are the best, which can achieve accuracy 0.778±0.028, sensitivity 0.779±0.092, specificity 0.778±0.030, and AUC 0.8284. Using Integrated gradients to visualize deep learning models, model adding Squeeze and Excitation block is likely to focus on the entire VOI. Model adding Dilate convolution block can focus on tumor margins, pleural tags, the junction of lung area and lung wall, etc. Model adding Lung map segmentation block can mainly focus on the whole tumor, tumor margins, pleural tags, the junction of tumor or lung area and lung wall, etc.
This study uses the deep learning network with Lung map segmentation block to test the external dataset, and compare it with the discriminant results of clinicians. The discriminant results of clinicians tend to be higher in sensitivity, but lower in accuracy and specificity. The deep learning model is tested 30 times, which can achieve accuracy 0.822±0.030, sensitivity 0.784±0.052, specificity 0.828±0.037. Although sensitivity is slightly lower than that of clinicians, both accuracy and specificity are much higher.
In this study, the external dataset is further divided into whether the tumor is in contact with the pleural. In the cases where the tumor has contact with the pleura, although the accuracy of the deep learning model in diagnosing visceral pleura invasion is slightly lower than that of clinicians, the accuracy is higher in diagnosing the absence of visceral pleural invasion. In addition, the results of clinicians fall within the 95% confidence interval of ROC curve of the deep learning model. The experimental results show that the deep leaning model discriminating visceral pleural invasion in this study can achieve similar results to that of clinicians.

並列關鍵字

Classification of visceral pleural invasion of lung tumors ； deep learning ； convolutional neural network ； attention mechanism ； visualization of deep learning model

參考文獻

[1] 108年國人死因統計結果-衛生福利部。民國108年06月16日。檢自https://www.mohw.gov.tw/cp-4631-54482-1.html

[2] Siegel, Rebecca L., Kimberly D. Miller, and Ahmedin Jemal. "Cancer statistics, 2019." CA: a cancer journal for clinicians 69.1 (2019): 7-34.

[3] Detterbeck, Frank C., et al. "The IASLC Lung Cancer Staging Project: methodology and validation used in the development of proposals for revision of the stage classification of NSCLC in the forthcoming (eighth) edition of the TNM classification of lung cancer." Journal of thoracic oncology 11.9 (2016): 1433-1446.

[4] Goldstraw, Peter, et al. "The IASLC lung cancer staging project: proposals for revision of the TNM stage groupings in the forthcoming (eighth) edition of the TNM classification for lung cancer." Journal of Thoracic Oncology 11.1 (2016): 39-51.

[5] Brindle, Lucy, et al. "Eliciting symptoms interpreted as normal by patients with early-stage lung cancer: could GP elicitation of normalised symptoms reduce delay in diagnosis? Cross-sectional interview study." BMJ open 2.6 (2012).